Deep Lake Explorer: A web application for crowdsourcing the classification of benthic underwater video from the Laurentian Great Lakes

Abstract

Underwater video is increasingly used to study aspects of the Great Lakes benthos including the abundance of round goby and dreissenid mussels. The introduction of these two species have resulted in major ecological shifts in the Great Lakes, but the species and their impacts have heretofore been underassessed due to limitations of monitoring methods. Underwater video (UVID) can "sample" hard bottom sites where grab samplers cannot. Efficient use of UVID data requires affordable and accurate classification and analysis tools. Deep Lake Explorer (DLE) is a web application developed to support crowdsourced classification of UVID collected in the Great Lakes. Volunteers (i.e., the crowd) used DLE to classify 199 videos collected in the Niagara River, Lake Huron, and Lake Ontario for the following attributes: round goby, Dreissena, and aquatic vegetation presence, and dominant substrate type. We compared DLE classification results to expert classification of the same videos to evaluate accuracy. Depending on the attribute, DLE had 77% (hard substrate) to 90% (vegetation presence) agreement with expert classification of videos. Detection rates, or the number of videos with an attribute detected by both volunteers and an expert divided by the number where only the expert detected the attribute, ranged from 62% (hard substrate) to 95% (soft substrate) depending on the attribute. Many factors affected accuracy, including video quality in the application, video processing, abundance of species of interest, volunteer experience, and task complexity. Crowdsourcing tools like DLE can increase timeliness and decrease costs but may come with tradeoffs in accuracy and completeness.

Predicting dreissenid mussel abundance in nearshore waters using underwater imagery and deep learning

“…Mussel analyses have been conducted using a variety of methods. Most recently, presence of invasive benthic species in short video clips was classified (Wick et al 2020). The method relies on a manual binary presence/absence determination made by crowdsourced volunteer labor and expert scientists.…”

“…On the other hand, volunteer experience is challenging to quantify, and the average experience level may not increase substantially over time due to turnover. Experience identifying invasive mussels is critical for accurate human classifications (Wick et al 2020). As a result, the scalability of volunteer labor is limited due to the limited experience of volunteers.…”